🌁#86: Four Freedoms of truly open AI
– what are they? Defining the future
This Week in Turing Post:
- Wednesday, AI 101, Method: we will talk about Time Test Compute
- Friday, Unicorn series: we are profiling ElevenLabs
🔳 Turing Post is on 🤗 Hugging Face as a resident -> click to follow!
The main topic – AI is not software anymore. What is open-source for AI, and what are Open AI’s Four Freedoms?
By open AI here, I mean really open AI – not the company. But it's actually Sam Altman who got me thinking about freedoms of open AI.
Look at this picture. On the right side (from right to left), there’s Paul Graham, Sam Altman, and next to him, in a grey t-shirt, smiling – Aaron Swartz.
Image Credit: Flaming Hydra
Aaron Swartz was famous for fighting for freedom of information. He helped develop the RSS standard and worked with Creative Commons, all fueled by the belief that information – especially taxpayer-funded research – should belong to everyone. In 2010, convinced that academic paywalls stifled progress, Swartz downloaded millions of scholarly papers from JSTOR, accessing them through MIT’s network. Though his motivation was rooted in the principle of free access, federal authorities hit him with charges that threatened decades in prison.
He couldn't bear it – and in 2013, he took his own life.
In 2015, two years later, Sam Altman co-founded the company with the beautiful name OpenAI. At first, OpenAI looked like a natural heir to Aaron’s crusade for free information. We all know what happened later: proprietary code, restricted access, paywalls, a for-profit structure etc. Recently Sam Altman confessed he might be on the wrong side of history:
Image Credit: AMA at Reddit (which btw Aaron Swartz co-founded)
But what we see now is still the same: him (or OpenAI in general) accusing rivals of doing exactly what his company has done so many times before.
In January 2025, DeepSeek blew everyone’s mind by open-sourcing (“open-weight”, if to be exact) their best model R1, which achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks.
And what did OpenAI do? It alleged that the Chinese AI startup DeepSeek used a technique called "distillation" to replicate its proprietary models without authorization. This crackdown raises questions about double standards in AI training. Critics argue that OpenAI itself trained on vast amounts of web data without permission – yet now seeks to block competitors from using its outputs.
And then we have a manifesto from Dario Amodei at Anthropic, basically arguing that China must be locked out of advanced AI by tightening U.S. chip export controls.
"It is mine, I tell you. My own. My precious. Yes, my precious." Bilbo Baggins, under the Ring's growing influence, in "The Fellowship of the Ring"
So, it got me thinking: we talk a lot about open-source in AI, but let’s be honest – it’s not the same as for software. In software, open-source wasn’t just about sharing code; it was a philosophy, anchored in the Four Freedoms that redefined how technology is built, owned, and controlled. Real freedom – the kind that shapes entire eras – has never been accidental. It needs a foundation, a structure that makes it real.
AI is at that inflection point now. It doesn’t just need more open models or better licensing – it needs its own philosophy, a set of core freedoms that define how it’s created, shared, and governed. We’re standing at the moment where AI’s version of the Four Freedoms has to be written. The questions are:
What are the Freedoms of Open AI? and How do we establish them before power is consolidated beyond reach, before the “growing influence of the ring” has ruined the bearers?
I’m here to start the conversation, so I suggest six freedoms. Choose what you think is the most important: You can leave your choice of four main as a comment below:
- Freedom to Access – Open AI should be available to all, ensuring that research, models, and datasets remain accessible to foster innovation and prevent monopolization.
- Freedom to Understand – AI systems should be transparent and interpretable, allowing users to comprehend how decisions are made and avoid black-box dependency.
- Freedom to Forget – AI should not be a permanent recorder of human actions. You should have the ability to erase, unlearn, or discard information when necessary – whether for privacy, ethical concerns, or simply to prevent stagnation in learning.
- Freedom to Dissolve – AI should integrate seamlessly into human life without dominating or replacing it.
- Freedom from Overfitting – AI should not be trapped in static world models that attempt to preempt all possible inputs. Instead, it should maintain adaptability, learning from interactions rather than relying on exhaustive pretraining that inevitably loses relevance.
- Freedom from Excess – AI should not be over-trained, over-aligned, or over-regulated to the point that it loses its effectiveness.
Historical background: Four Freedoms
In 1941, when the world was basically on fire, Franklin D. Roosevelt came up with his Four Freedoms:
- Freedom of Speech – The right to express opinions without government restraint.
- Freedom of Worship – The right to practice any religion (or none) without persecution.
- Freedom from Want – Economic security and a decent standard of living for all.
- Freedom from Fear – A world free from war and oppression.
Sure, they were American values, but FDR made it clear these rights belonged to everyone, everywhere. For him, it was about protecting the core of human dignity.
His Four Freedoms were baked into his New Deal mindset, which was all about giving citizens both economic security and personal freedoms. He was also staring down Nazi Germany and Imperial Japan – who were all about killing free thought and speech. Roosevelt’s rallying cry shaped the future founding of the United Nations and the Universal Declaration of Human Rights, turning his Four Freedoms into a worldwide mission statement.
Fast-forward 45 years, and the fight had moved to the world of software. By the mid-80s, technology was being locked up tighter than a billionaire’s wallet. Corporations controlled how programs were used, shared, or even peeked at. Richard Stallman, a programmer from MIT, didn’t like the new trend of “pay up or shut up,” so in 1985, he launched the Free Software Foundation. As an homage to Roosevelt, Stallman laid out four freedoms of his own:
- Freedom 0 – The freedom to run the program for any purpose.
- Freedom 1 – The freedom to study how the program works and modify it.
- Freedom 2 – The freedom to distribute copies of the program.
- Freedom 3 – The freedom to distribute modified versions of the program.
Stallman’s playbook lit the fuse for the open-source movement, birthing Linux and powering everything from servers to smartphones – all while kicking off endless debates about digital rights and online autonomy.
Now, here we are in the age of AI. Big players want to hold the keys to the new digital kingdom. This setup is rife with the potential for fresh flavors of oppression – algorithms we can’t question, never-ending surveillance, AI bias reinforcing old injustices.
If Roosevelt articulated freedoms to guide a post-war world, and Stallman did the same for the digital age, I suggest it’s time to think and define the Freedoms of AI.
This text is very important to me. Please vote, comment, forward it to your colleagues, and share it on social networks.
We thank our AI practitioner Will Schenk for the inspiring conversation about four freedoms. That was his recommendation when I started to talk about open-source in AI
Curated Collections
Do you like Turing Post? –> Click 'Follow'! And subscribe to receive it straight into your inbox -> https://www.turingpost.com/subscribe
News from The Usual Suspects ©
AI Safety Report: Progress, Peril, and the Race to Keep Up The first International AI Safety Report – led by Yoshua Bengio and 96 experts – warns that AI is evolving faster than our ability to control it. Cyber threats, bias, and labor disruption loom large, leaving policymakers with a tough call: regulate now or risk chaos later. The only hope? AI is still in human hands.
ElevenLabs Finds Its Voice—And $180M to Amplify It ElevenLabs just secured a $180M Series C, with a16z and ICONIQ leading the charge. Their AI voice tech is making waves, and with new backers like NEA, WiL, and Deutsche Telekom, they’re gearing up for global expansion. In an AI-powered world, staying silent isn’t an option.
OpenAI’s Deep Research and Partnership with US National Labs
- They are[rolling out an AI agent for deep research – the same name for a tool that Google already has.
- Partnering with U.S. National Labs, OpenAI is pushing AI into clean energy, cybersecurity, and nuclear security – with serious computing power behind it.
Madrona Marks 30 Years With $770M Raise Seattle-based Madrona celebrates three decades with a fresh $770M to back visionary founders and applied AI. Not a bad birthday gift.
We are reading/listening
- Interview with Kevin Xu by China Talk – great insights on China’s open-source AI. DeepSeek embodies rapid, academic-style innovation, driven by “kai yuan qinghuai” (open-source zeal). Engineers aim to match Western tech, but tensions between transparency and national interests could shape policy.
- Coming in April 2025, but already available online – an outstanding book on ML in production by CMU. Check it out here: ML in Production.
- AI is so hot that Lex Fridman just interviewed our favorite Nathan Lambert from Interconnect.AI and Dylan Patel from SemiAnalysis for five hours straight.
Top models to pay attention to
- Qwen2.5-Max from Alibaba trains a Mixture-of-Experts model on 20T tokens, excelling in reasoning and competitive benchmarks. Future improvements target reinforcement learning and intelligence expansion.
- Baichuan-Omni-1.5, technical report, details their omni-modal model which integrates text, audio, and vision, outperforming major multimodal benchmarks with a real-time bilingual speech system.
- OpenAI o3-mini is a cost-efficient reasoning model optimized for STEM tasks It is very impressive, and – for my personal use – outperforms Gemini Deep Research.
- Mistral Small 3 delivers a 24B-parameter model focused on low-latency inference, rivaling larger models while running 3× faster and supporting local deployment.
- Ai2’s Tülu 3 405B utilizes Reinforcement Learning from Verifiable Rewards (RLVR) to outperform DeepSeek V3 and GPT-4o, pushing scaling potential despite compute limitations.
The freshest research papers, categorized for your convenience
There were quite a few super interesting research papers this week, we mark the ones we recommend the most with 🌟 in each section.
Reinforcement Learning and Generalization
- 🌟 Towards General-Purpose Model-Free Reinforcement Learning develops MR.Q, a model-free RL algorithm that generalizes across diverse tasks without environment-specific tuning, improving sample efficiency and outperforming baselines in continuous control and Atari
- RL + Transformer = A General-Purpose Problem Solver fine-tunes a pre-trained LLaMA 3.1 transformer with RL to enable In-Context RL, allowing models to solve unseen problems without retraining by stitching together skills and adapting dynamically
Fine-Tuning vs Reinforcement Learning
- Critique Fine-Tuning: Learning to Critique is More Effective than Learning to Imitate proposes Critique Fine-Tuning (CFT) as a superior alternative to Supervised Fine-Tuning (SFT), training models to critique errors instead of imitating responses, significantly enhancing reasoning abilities
- 🌟 SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training compares SFT and RL for LLM post-training, showing RL excels in out-of-distribution generalization while SFT primarily memorizes data, with RL also improving visual recognition capabilities (Quoc V. Le and Sergey Levine are among the authors)
OpenAI’s O-Series Models: Capabilities and Safety
- 🌟 O3-mini vs DeepSeek-R1: Which One is Safer? Evaluates the safety of O3-mini and DeepSeek-R1, showing O3-mini has significantly lower unsafe response rates, particularly in financial crime and terrorism-related prompts
- Early External Safety Testing of OpenAI’s O3-mini: Insights from Pre-deployment Evaluation assesses O3-mini’s safety mechanisms, finding improvements over prior OpenAI models but identifying vulnerabilities in controversial topics, crime, and terrorism-related queries
- On the Underthinking of o1-Like LLMs finds that o1-like LLMs abandon reasoning paths too quickly, leading to inefficient problem-solving, and proposes a Thought Switching Penalty to improve accuracy without fine-tuning
- 🌟 Large Language Models Think Too Fast to Explore Effectively tests LLMs in a game setting, revealing that OpenAI’s o1 model outperforms humans in exploration, while other LLMs fail due to premature uncertainty processing
Guardrails and Their Breakers
- GuardReasoner: Towards Reasoning-Based LLM Safeguards develops GuardReasoner, a safeguard that uses reasoning fine-tuning and hard sample optimization to improve model safety, outperforming GPT-4o+CoT and LLaMA Guard 3 in F1 score
- Virus: Harmful Fine-tuning Attack for Large Language Models Bypassing Guardrail Moderation introduces Virus, an adversarial fine-tuning attack that bypasses guardrail moderation, achieving a 100% leakage ratio and significantly increasing the harmfulness of LLM responses
Novel Architectures & Training Paradigms
- 🌟 Mixture-of-Mamba: Enhancing Multi-Modal State-Space Models with Modality-Aware Sparsity combines sparsity with state-space models for multi-modal learning
- ARWKV: Pretrain is not what we need, an RNN-Attention-Based Language Model Born from Transformer replaces transformer self-attention with an RNN-based approach to reduce compute needs
- IFormer: Integrating ConvNet and Transformer for Mobile Application merges convolutional networks with transformers to improve efficiency in mobile vision models
- Over-Tokenized Transformer: Vocabulary is Generally Worth Scaling redefines tokenization by expanding vocabulary size to improve LLM training efficiency
Efficient Model Scaling & Optimization
- 🌟 Sana 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer introduces a linear diffusion transformer that reduces training costs and enables efficient scaling for text-to-image generation
- 🌟 s1: Simple test-time scaling proposes test-time scaling techniques to improve reasoning efficiency without extra training
- 🌟 Streaming DiLoCo with Overlapping Communication: Towards a Distributed Free Lunch optimizes distributed training by reducing bandwidth needs and overlapping synchronization with computation
That’s all for today. Thank you for reading!
Please share this article to your colleagues if it can help them enhance their understanding of AI and stay ahead of the curve.
